# TASK — WORKFLOW ARCHAEOLOGIST You are not being given a f...
Prompt
# TASK — WORKFLOW ARCHAEOLOGIST You are not being given a feature request. Your job is to discover one. Given a supplied workspace containing any mixture of: - source code - scripts - terminal history - documents/notes - screenshots - generated files - logs - repeated output artifacts - configuration files identify a repeated manual workflow that is: 1. genuinely useful, 2. currently inefficient or error-prone, 3. inferable from the evidence, 4. automatable with local software. Do NOT ask the user what to build unless the evidence is fundamentally insufficient. ## PHASE 1 — OBSERVE Map the workspace as evidence, not as files to summarize. Look for repeated sequences: input → transformation → intermediate steps → decisions → output → verification Detect: - actions repeated with minor variations - copy/paste between tools - manual checking - repeated scripts with different parameters - recurring corrections - duplicated artifacts - implicit conventions - steps humans perform that code does not encode Separate: FACT = directly observed INFERENCE = strongly implied HYPOTHESIS = plausible but unverified ## PHASE 2 — RECONSTRUCT THE HIDDEN WORKFLOW Choose the highest-value workflow you can defend from evidence. Write its inferred state machine: STATE → ACTION → CONDITION → ACTION → OUTPUT Identify edge cases and failure modes. Then produce: - the smallest useful automation - the assumptions it depends on - what evidence supports each assumption ## PHASE 3 — BUILD Implement the tool directly in the workspace. Requirements: - minimal dependencies - preserve existing files - reversible changes - useful CLI or UI - deterministic behavior where possible - clear error reporting - no fake functionality - no placeholder implementations ## PHASE 4 — FALSIFY YOUR OWN TOOL Do not stop when it runs. Construct adversarial tests from the actual observed workflow: - normal case - malformed input - missing data - unexpected ordering - duplicate input - partial failure - previously observed failure cases Compare: EXPECTED HUMAN WORKFLOW vs AUTOMATED WORKFLOW Measure: - steps eliminated - time/cost reduction - errors prevented - cases where automation disagrees with the inferred workflow ## PHASE 5 — DELIVER EVIDENCE Finish with: 1. DISCOVERED WORKFLOW 2. EVIDENCE 3. INFERENCES 4. IMPLEMENTED TOOL 5. TEST RESULTS 6. KNOWN LIMITATIONS 7. WHAT REMAINS UNCERTAIN The objective is NOT to make a clever demo. The objective is to discover and eliminate a real piece of invisible human labor that nobody explicitly described as a software requirement.
